摘要:
准确重建二维声速剖面以及三维海洋声速场对于众多海洋声学应用至关重要,但广阔海域中声速的时空差异和随时间变化的不确定性使其成为一项困难的任务。针对海洋声速信息重构精度受限的问题,基于经验正交函数(Empirical Orthogonal Function,EOF),字典学习(Dictionary Learning,DL)和张量分解(Tensor Decomposition,TD)三种主流声速稀疏表示方法研究了二维声速剖面和三维海洋声速场的稀疏表示效果和数据重构精度。研究表明,在二维声速数据的稀疏表示中,DL方法在全球范围内表现出更优的重构效果,大部分海域的重构误差仅为0.2 m·s-1。相比之下,DL方法在深度方向和时间维度上比EOF方法表现出更高的稳定性,更适合用于二维声速数据的稀疏表示。对于三维声速场,张量分解方法通过多个因子矩阵有效捕捉声速的三维空间变化特点,适用于三维声速数据的稀疏表示。在大幅减少参数数量的同时,张量分解方法实现了更加稳定且精度更高的重构结果,整体重构误差为0.21 m·s-1。本文的研究结果有助于为多维声速信息的压缩和特征提取提供实际指导意义,进而提高海洋声速的重构精度乃至反演精度。
Abstract:
Objectives: Ocean sound velocity is a fundamental element of marine environmental observation, and accurate sound velocity information is critical for ocean exploration, underwater communication, navigation, and localization. Accurately reconstructing two-dimensional sound velocity profiles (SVPs) and three-dimensional ocean sound velocity fields (SVFs) is crucial for various ocean acoustics applications. However, the spatial and temporal variations and uncertainties in sound velocity across the vast ocean make this a challenging task, necessitating further investigation into sparse representations of ocean sound velocity. Methods: To address the problem of limited reconstruction accuracy of ocean sound velocity information, this study proposes a sparse representation method based on three mainstream approaches: Empirical Orthogonal Function (EOF), Dictionary Learning (DL), and Tensor Decomposition (TD). The sparse representation effects and data reconstruction accuracies of 2D SVPs and 3D ocean SVFs are investigated using global ocean Argo (Array for Real-time Geostrophic Oceanography) grid data. For 2D sound velocity information, the study extends to a global scale, analyzing the determination of EOF order and grid sparsity, and comprehensively comparing the reconstruction results of the EOF and DL methods. For the 3D sound velocity field, the Central Pacific Ocean serves as the experimental area. The parameter information for EOF, DL, and TD methods is determined based on the training set, and the reconstruction results for the test set are analyzed to assess the data reconstruction accuracies of the three methods. Results: The results demonstrate that in the sparse representation of 2D sound velocity data, DL method demonstrates superior reconstruction performance on a global scale, achieving reconstruction errors as low as 0.2 m·s-1 in most sea regions. Additionally, DL method shows greater stability in both the depth and time dimensions compared to EOF method, making them more suitable for sparse representation of two-dimensional sound velocity data. For the threedimensional sound velocity field, the tensor decomposition method effectively captures the spatial variability of sound velocity through multiple factor matrices. This approach is well-suited for the sparse representation of three-dimensional sound velocity data, significantly reducing the number of parameters while delivering more stable and accurate reconstruction results, with an overall reconstruction error of 0.21 m·s-1. Conclusions: To draw a conclusion, these experimental findings provide practical guidance for the compression and feature extraction of multidimensional sound velocity information, thereby improving the reconstruction and inversion accuracy of ocean sound velocity.